Assessing the Effectiveness of Transfer Learning in Chest X-Ray Analysis With the BRAX Dataset

The integration of machine learning techniques in medicine has significantly enhanced medical diagnostics, making them more efficient, agile, and accurate. With the continuous evolution of medical datasets, adapting machine learning models to distinct clinical contexts – often characterized by domain shift and data scarcity – is crucial. In this scenario, transfer learning and fine-tuning strategies play a key role in optimizing the training by adapting pre-trained models to new data effectively. This study investigates the impact of different fine-tuning strategies on the classification of pathologies in chest X-ray images. Instead of focusing solely on model comparisons, we analyze how various fine-tuning approaches affect performance when transferring knowledge from the CheXpert dataset to BRAX (Brazilian Labeled Chest X-ray dataset). The architectures DenseNet-121, ResNet-101, and Swin Transformer are employed as tools to assess the effectiveness of these strategies. The fine-tuning techniques explored include full layer freezing, gradual unfreezing of layers, and L1-SP/L2-SP regularization. Model performance is evaluated using ROC AUC, precision, recall, and F1-score, with 5-fold cross-validation ensuring result validity and generalization. Our findings demonstrate that the choice of fine-tuning strategy significantly impacts performance, surpassing the baseline where models were pre-trained on ImageNet and trained exclusively on BRAX. These results emphasize the importance of tailored fine-tuning approaches for optimizing deep learning models in medical imaging.